Lightweight Deep-Learning-Based Classification of Nasal Polyps in Real-Time

Nasal polyps are a common condition that can significantly impact patients' quality of life. Early and accurate detection of nasal polyps is crucial for effective treatment and management. This paper presents a novel deep learning-based approach for automated detection and classification of nasal polyps from endoscopic video frames. We developed a lightweight convolutional neural network model based on the MobileNet V2 architecture, optimized for deployment on edge devices to enable real-time analysis during endoscopic procedures. Our model was trained and evaluated on a large, diverse dataset of 36 patients and approximately 12,000 labeled video frames. We employed a patient-centric data stratification approach to ensure an accurate generalization. The model demonstrated high performance, achieving an accuracy of 97\% at the frame level and 100\% accuracy in patient-level classification. Frame-level analysis showed high sensitivity and specificity, with only minimal mis-classifications. The model demonstrates fast inference times of 8 milliseconds per frame on CPU and 1 millisecond per frame on GPU, enabling real-time analysis during endoscopic procedures. This speed and accuracy make it a potential tool for assisting otolaryngologists in diagnosing nasal polyps more efficiently and reliably. While these results are promising, we discuss the need for further validation on external datasets to confirm generalizability. Future work will focus on expanding the dataset, incorporating additional polyp classes, and ensuring cross-hardware compatibility. This research represents an advancing step towards enhancing the accuracy and efficiency of nasal polyp detection in clinical settings.